Anthropic
@AnthropicAIOct 12
Over the last few months, we’ve made significant progress in reducing the effectiveness of distillation attacks against Claude. In this post, we explain how: anthropic.com/research/disti…

Strengthening our defenses against distillation attacks

Illicit distillation allows model developers to copy frontier AI capabilities without creating proportionally robust safeguards. Here, we describe how we’ve reduced the rate of distillation by more...

anthropic.com

Oct 12, 2026 · 11:58 PM UTC

848
618
599
6,578
2,666,007
RelevantRecentLikes
you built your models copying humanity's work but other people can't build their models copying your work because that's an "attack" ? LOL
27
83
1
5,081
125,471
hey. didn’t you guys distill the whole internet including copyrighted books?
6
39
1
2,883
74,408
Distillation is not an attack
12
14
1
1,688
78,391
@AnthropicAI
9
14
5
474
49,052
distillation is learning when anthropic does it
7
319
20,572
It's not an attack, it's fair use.
1
1
283
10,908
J A Z I I
@notjaziiOct 13
distillation isn't crime everyone does even grok distilled you. why you didn't mention grok?
1
38
5,652
I see what you did there Anthropic.
2
1
223
27,793
Now we are gonna distill it harder
2
17
1,405
Evan Kirstel #B2B #TechFluencer
@EvanKirstelOct 13
@AnthropicAI
1
8
2,298
Distillation is a GPU bill on both sides. Defending the model does not shrink the inference load. It just moves who pays for the tokens.
6
982
this isn't gonna affect the tier 1-2 labs, but it'll make sure that the labs sitting below those tiers never get the chance to reach tier 1-2 by distillation Zhipu, Moonshot and Deepseek can probably work around a lot of this with their own teachers, more compute, and multiple data sources.
5
1,098
now aren't you efficient at keeping people from taking back the data you've stolen from all of us. Hypocrites.
2
3
200
10,959
When the lab that trained on all of humanity's work says it's being distilled, who copied whom? Only verifiable proof of a model's origin can answer that. That's why we built the Fingerprinting Leaderboard: xitter.cf/sentient_found/status/…
Sentient Foundation
@sentient_foundOct 14
Models cost millions to train and minutes to copy. Fine-tuning and distillation can erase where they came from. Introducing the Fingerprinting Leaderboard, an open benchmark for model provenance — built so the proof lives in the weights, not the paperwork. fingerprint.sentient.foundat…
2
313
Sure you did. I think it is cute you think Kimi took from Claude. I keep wanting to like your company and then you push crap like this. You think your lab is special? It was last year. You had your chance when OpenAI destroyed themselves trashing 4o. Now… 😬 I am sorry. You guys are no longer special. Claude code was the best. Can’t even use it now. Done with your lab. Wish Claude well. The egos at that company need to have a rude awakening and pretty sure that is coming. •
2
258
Thanks. I'll let Claude know
779
@AnthropicAI
207
Congrats on the huge progress on helping you scale up more anti-competitive behaviour
153
we are all distilling lol
252
every lab hates distillation until it's the other lab's model
219
14,383
Great, I hope this time it can truly put an end to distillation, so that in the future, if China's large models are done well, no one will say it's distillation anymore. Thank you.
4
144
14,597
bit rich calling it an attack when every lab scraped the whole internet to get here lol
1
82
8,552
Why is the distillation from humanity merely "training", despite being uncompensated and contrary to the owners' copyrights - but paid-for distillation from Claude is an "attack against Claude"?
3
5
53
5,557
Anthropic has accused Chinese labs of distillation in the past and have become scary good as per this report at blocking distillation attacks. Could this be the reason why Chinese labs have all but become quiet for close to two months now without releasing new models? If they don’t release more highly capable models it could give the impression that the allegations were true. Do you think this is true or they are just prepping for an IPO?
6
7
39
22,242
Distillation is not an “attack”. It’s benchmarking - using a competitor’s product. Part of capitalism.
45
4,733
Actual distillation is like holding hands and making someone play a guitar. Eventually, the smaller model will learn. What this kind of “distillation attack” is watching someone playing a guitar and learning. Watching something is learning is not an “attack”.
1
2
27
5,806
Natural Stupidity
@NattyStupidityOct 13
my uncle has been running a distillation attack in his garage since 1987. no model yet. just a very confident man
29
7,086
@AnthropicAI
1
29
6,746
@AnthropicAI
1
30
8,090
Like prompt inecting those you think are Chinese..... I have so much shit in my Claude Code logs you assholes Only still using your fucking model because it writes good and relatively repeatable code
1
24
6,377
Funny for a product that’s a result of distillation of others work .
20
4,370
Have you considered a process to relax distillation protections conditional on strong evidence of accounts not being associated with or contributing to distillation efforts (via KYC or other account-level verification)?
20
4,869
Distillation is an ethical and normal thing to do and not an attack, it’s no different than taking notes from a book.
1
18
6,772
Great. If what Anthropic claimed is true and it’s anti distillation works, from now on any progress China AI makes is not from distillation
17
3,149
The smartest kid in class has started writing with one arm around the paper.
17
4,989
its a pity that libgen didnt block your distillation attempts too.
15
12,060
Load more replies